Elon Salfati

Where are you from, and what is your background?

I’m from Israel and now live in Zürich. I came to AI through security. My first real job was in counter-intelligence, and it taught me to look at any system and ask how it could break. I’ve kept asking that question ever since, first in security and then in machine learning and AI. Along the way I’ve started three companies, because I’ve always preferred building things myself to watching from the side. Today I split my time between my company and my PhD. Each one feeds the other: the problems I see in industry shape my research, and the research changes how I approach problems.

What do you do in your spare time?

Honestly, I don’t have much of it. When I do take time off, I make the most of living in Switzerland, which mostly means skiing in winter. I also spend a lot of time going back and forth between Zürich and London, and I’m slowly learning German. The rest of the time I’m thinking about what the next big problem worth solving is. That’s probably not a hobby, but it’s what I enjoy most.

What influenced you to do a PhD?

I was never very academic. But over the last few years my work in industry kept pulling me toward open research questions, and my mix of intelligence, security, enterprise software and AI gave me an unusual view of them. A PhD at SAIL was how I could go further than industry lets you. You don’t know what you don’t know, and the PhD has already shown me a lot: new ways to do research, new fields, and new ways to understand what happens inside large neural networks. It has also shown me how easy it is to make an AI system unsafe, and that is a big part of why I think this work matters.

What are your research interests?

I study how AI systems can be attacked and how to defend them. I don’t think you can secure a model only from the outside, with filters on what goes in and what comes out, because many weaknesses are created during training and live inside the model’s weights. So my work sits where AI security meets model training, interpretability and alignment. I want to understand how a model learns unsafe behavior, how to find that behavior inside the network, and how to remove it instead of hiding it. The goal is AI systems that stay safe even when someone is actively trying to break them.

← All people